Abstract

Lending environments have become more complex, and the frequency of loan defaults has increased, making it a need for intelligent, data-driven approaches for borrower evaluation. Traditional credit scoring approaches do not fully account for the credit risk of borrowers from a multi-dimensional perspective, such as financial, behavioral, and non-financial factors like ESG metrics. Understanding these trends, this study introduces an innovative multi-stage AI-based model named IBP-RLDPO, designed to enhance borrower profiling, risk-aware lending decisions, and portfolio optimization by using LightGBM, ensemble learning, feature engineering, and explainable AI. The framework integrates multi-source data, pre-processes the data, generates borrower risk profiles based on PD, LGD, and EAD models, and uses a risk-aware decision engine with portfolio-level optimization. Using SHAP and LIME for explainability layers can offer insights into the contributions of various features, which helps in facilitating transparent and ethical lending practices. On the Lending Club dataset of 887,379 loan records, an empirical assessment shows that it outperforms baseline and ensemble models with 98.25% accuracy, 78.5% recall, 98.25% precision, 82.5% F1-score, and 94.25% AUC. In a nutshell, IBP-RLDPO can improve risk adjudication of the decision-making process, help manage regulatory compliance, mitigate the risk of bad loans, and maximize returns on good loans. This research provides a full-scale and interpretable solution that is scalable for intelligent lending and optimized portfolio management, providing measurable business value and operational efficiency.